Schwarz A.L. Machine Learning for Human Motion Analysis and Gesture Recognition

Technical University of Munich, 2012. — 163 p.In this thesis, we investigate machine learning methods for human motion analysis. We introduce algorithms for human pose estimation and activity recognition that do not rely on classical cameras and that can cope with noisy and incomplete input data. We propose methods that capture human movements using body-worn inertial sensors or using a depth camera. In a training phase, the measurements from these modalities are complemented with precise motion data recorded with a camera-based system. We present a novel technique for learning human motion models from this data using non-linear dimesionality reduction and regression, enabling human pose estimation and activity recognition given only the inertial sensor or depth data. As an application scenario, we present a customizable method for gesture-based interaction in the operating room that fills an existing gap, as confirmed by surgeons.